---
title: "Strong planner, cheap workers: one strong model coordinating many cheap ones"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Strong planner, cheap workers: one strong model coordinating many cheap ones}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include=FALSE}
knitr::opts_chunk$set(
  collapse = TRUE, comment = "#>",
  eval = identical(tolower(Sys.getenv("LLMRAGENT_RUN_VIGNETTES", "false")), "true")
)
```

Strong models cost the most on the task where cheap models are weakest: broad
exploration that then has to be synthesized. Cheap models produce many
independent drafts but combine them poorly. `think_harder()` uses this
division of labor:

1. **Plan.** The strong model decomposes the problem into genuinely different
   lines of attack (one call).
2. **Work.** A cheap model pursues each line independently and in parallel,
   blind to the others, so the drafts do not converge by imitation.
3. **Synthesize.** The strong model weighs the drafts against one another,
   discards what fails scrutiny, and writes the answer (one call).
4. **Verify.** Optionally, the strong model attacks its own synthesis as a
   hostile reviewer; one revision pass runs when it finds a substantive flaw
   (one or two calls).

Whatever the fan-out width, the strong model is billed for two to four calls;
the volume goes to the cheap model. And unlike a single unsupported answer,
every intermediate product is kept for inspection.

```{r run}
library(LLMRagent)

strong <- LLMR::llm_config("deepseek", "deepseek-reasoner")
cheap  <- LLMR::llm_config("groq", "openai/gpt-oss-20b", temperature = 0.8)

out <- think_harder(
  "A mid-sized university wants to raise its course-evaluation response rate
   from 35% to 70% within two semesters, without making responses mandatory
   and without raffles or payments. Design the most promising intervention
   portfolio, with predicted effect sizes where evidence exists.",
  strong_config = strong,
  cheap_config  = cheap,
  n_approaches  = 5
)

cat(out$answer)
```

The audit trail:

```{r inspect}
out$plan                                  # what the planner asked for
out$workers[, c("approach", "success")]   # who delivered
out$verification                          # what the reviewer objected to
out$revised                               # whether a repair pass ran
LLMR::llm_usage(out$workers)              # tokens spent on the cheap side
```

Two practical notes. First, worker drafts are independent by construction;
if you see near-identical drafts, your approaches were paraphrases, and the
remedy is a better problem statement, not more workers. Second, the pattern
composes with `LLMR::llm_log_enable()`: turn it on beforehand and the whole
orchestration, every worker included, is written to one auditable JSONL file.

`think_harder()` is a fixed pipeline: plan, work, synthesize, verify. When
you want the strong model to decide for itself when and whom to consult,
build the same arrangement directly with `agent_as_tool()`: give a strong
supervisor several cheap specialists as tools and let it route the work.